Coding Horizon

Letta Makes Local AI Finally Remember Your Work

Every figure, name and claim the finished picture puts on screen, chased to a primary source. Checked 7 September 2026.

Letta on GitHub

The narration says Letta has more than twenty five thousand GitHub stars across its repos. Measured from the GitHub REST API on 7 September 2026, across the 61 public repositories in the letta-ai organisation:

Repository	Stars
letta-ai/letta	24,645
letta-ai/letta-code	3,224
letta-ai/claude-subconscious	2,885
letta-ai/agent-file	1,197
letta-ai/lettabot	326
letta-ai/trajectory	247
Organisation total, 61 public repos	34,020

The main platform repository alone is 24,645, and it plus letta-code is 27,869, so “more than twenty five thousand across its repos” holds with room to spare.

Source: GitHub REST API, api.github.com/orgs/letta-ai/repos and api.github.com/repos/letta-ai/letta, read 7 September 2026.

What a Letta agent is given

Letta describes itself as a platform for stateful agents with memory that can learn and self improve over time. The agent surface includes persistent memory, identity, skills, tools, subagents and schedules.

• Skills are “directories containing instructions and resources that your agent can load when relevant”, scoped to the project, to the agent’s own memory, or bundled with the harness. Source: https://docs.letta.com/configuration/skills
• Subagents: “It is possible to deploy an arbitrary Letta Agent as a subagent by specifying its agent_id.” Source: https://docs.letta.com/letta-agent/subagents
• Schedules drive recurring work such as planning, triage and reporting. Source: https://docs.letta.com/handbook

MemFS, the git backed memory filesystem

Letta agents use MemFS, described in the documentation as “a git-backed memory filesystem that they can inspect and edit”. The memory itself “is part of the agent’s state, held in a git repository that belongs to the agent”.

• Projection: “Each memory is projected as Markdown with YAML frontmatter.”
• Directory shape, as published:

$MEMORY_DIR/
├── system/
│   ├── persona.md
│   └── human.md
├── reference/
│   └── project-notes.md
└── skills/
    └── my-skill/
        └── SKILL.md

• What is in context: “Files under system/ are loaded into the agent’s system prompt on every turn.” Everything else: “Files outside system/ stay out of context until they are needed.”
• Version history: “Every memory edit is committed to the MemFS git repository. This provides version history, conflict resolution, and a clear boundary between saved memory and uncommitted changes.”

Sources: https://docs.letta.com/concepts/memfs and https://docs.letta.com/configuration/memory

Models, and swapping them under a running agent

Supported hosted providers include Anthropic (Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, Claude Haiku 4.5), OpenAI (GPT 5.2, GPT 5.2 Codex) and Google (Gemini 3 Pro, Gemini 3 Flash).

Local inference providers are supported for local agents: Ollama, LM Studio and llama.cpp. The documentation states that direct integrations with local inference providers such as LM Studio, Ollama and llama.cpp are available for local agents, and that to keep inference local you connect one of them.

On switching: “Agents keep their memory and tools when you change models.”

Sources: https://docs.letta.com/v1-sdk/models, https://docs.letta.com/configuration/models and https://docs.letta.com/self-hosting

Context-Bench

Letta’s Context-Bench measures “how well language models can chain file operations, trace entity relationships, and manage multi-step information retrieval in long-horizon tasks”. Published 30 October 2025.

Model	Score	Benchmark cost
Claude Sonnet 4.5	74.0%	$24.58
GPT 5	72.67%	$43.56
GPT 5 mini	64.33%	$12.45
GLM 4.6	56.83%	not published in the post
Kimi K2	55.13%	$12.08
GPT 5 nano	44.83%	not published in the post
GPT OSS models	6.67% to 20.2%	not published in the post
DeepSeek V3	11.97%	not published in the post
GPT 4.1 nano	16.2%	not published in the post

GPT 5’s $43.56 against Claude Sonnet 4.5’s $24.58 is 1.77 times the cost for 1.33 points less, which is what the narration means by “almost twice the benchmark cost”. Kimi K2 at $12.08 for 55.13% is the best cost per point of the open weight entries.

Source: https://www.letta.com/blog/context-bench/ and the live board at https://leaderboard.letta.com/

Memory in production agents

Letta’s July 2026 evaluation covers adherence, retrieval, generalization and hygiene, across tasks that generate, clean and repair memory. Published 28 July 2026.

• “Anthropic models lead overall, with the widest margin in memory generation.”
• “OpenAI models offer strong cost-performance but trail on memory generation.” GPT 5.6 Sol leads among the OpenAI entries, and smaller models fall off more sharply when they have to rewrite memory rather than use existing context.
• Among open weight models, Kimi K3 is the strongest, and is better at repairing memory than at extracting a durable lesson, ahead of MiniMax M3 and GLM 5.2.

Source: https://www.letta.com/blog/evaluating-memory-in-production-agents

Memory rot

The term is Letta’s own. From “Memory Models: Towards Agents That Learn”, published 25 June 2026: “This can lead to memory rot over time for long-lived agents, and limit the extent to which agents can learn independently from available context or past experience.” The same piece identifies overgeneralization and memory rot as failure modes of existing memory-creation approaches. Hygiene is one of the four axes of the July 2026 production memory evaluation above.

Sources: https://www.letta.com/blog/towards-agents-that-learn/ and https://www.letta.com/blog/evaluating-memory-in-production-agents

The product surface

• Letta Code runs in the terminal, installed with npm install -g @letta-ai/letta-code and launched as letta. Node 22.19 or newer.
• The desktop app ships for macOS (Apple Silicon), Windows (x64 and ARM64) and Linux (AppImage, x64 and ARM64). Agents created in the CLI appear in the desktop app and the other way round.
• An App Server runs local or self hosted agents; Letta Cloud runs cloud agents whose state is available across machines.
• Channels carry messages in and out. The documented set is Slack, Telegram, Discord, WhatsApp and Signal, plus custom channels, which are headless and configured from local config files.

Sources: https://docs.letta.com/letta-agent/desktop-app, https://github.com/letta-ai/letta-code, https://docs.letta.com/self-hosting, https://docs.letta.com/configuration/channels and https://docs.letta.com/configuration/channels/custom

What actually fits in eight, sixteen and twenty four gigabytes

Download sizes for the default quantization of each tag, read from the Ollama registry manifests on 7 September 2026. These are weights on disk, which is the floor for what has to be resident; context and key value cache are on top.

Model	Size
Qwen2.5 Coder 7B	4.7 GB
Llama 3.1 8B	4.9 GB
Qwen3 8B	5.2 GB
Gemma 3 12B	8.1 GB
Qwen3 14B	9.3 GB
Qwen3 Coder 30B	18.6 GB

That is the shape the narration describes: a seven or eight billion parameter model clears eight gigabytes but leaves little for context, twelve to fourteen billion is comfortable at sixteen, and a thirty billion parameter coding model needs a twenty four gigabyte card before context is affordable.

The RTX 3090 and RTX 4090 both carry 24 GB of GDDR6X.

Sources: registry.ollama.ai/v2/library/<model>/manifests/<tag>, read 7 September 2026, and https://www.nvidia.com/en-us/geforce/graphics-cards/

Not chased to a primary source

• The narration lists “browser” among Letta’s channels. The documented channel set is Slack, Telegram, Discord, WhatsApp, Signal and custom channels; a browser channel is not named in the documentation, so it is not drawn on screen.
• Context-Bench costs for GLM 4.6, DeepSeek V3, GPT 5 nano and the GPT OSS entries are not published in the announcement post, so only their scores are drawn.